v0.3.0

v0.3.0: Final Submission

Final submission for DSCI 575. Scales the pipeline from the Milestone 2 state, adds an LLM comparison, and completes documentation and cloud-deployment planning.

Highlights

• Scale (Option 3): pipeline now samples 120,000 reviews and builds a 110,167-document corpus after preprocessing. BM25 and FAISS indexes rebuilt against the larger corpus.
• LLM experiment: LLaMA 3.3 70B vs LLaMA 3.1 8B on 5 queries with identical retrieval context and the same BALANCED prompt. LLaMA 3.3 70B selected as the default based on completeness, grounding, and filtering of off-topic retrievals.
• App: Streamlit app (app/app.py) supports BM25, Semantic, Hybrid, and RAG modes over the full corpus, with a Groq model fallback chain for resilience.
• Cloud deployment plan: S3 for storage, Elastic Beanstalk for compute, Groq managed API for inference, EventBridge + AWS Batch for nightly re-indexing. Architecture diagram and justification in results/final_discussion.md.

Docs and code quality

• README updated for the scaled pipeline, new active-retriever note, demo, and runtime/disk footprint.
• All public functions and classes in src/ have docstrings.
• No hardcoded paths, no API keys in source.
• environment.yml trimmed of unused packages.

Where to look

• results/final_discussion.md: full write-up: scaling, LLM comparison, cloud plan.
• notebooks/08_llm_comparison.ipynb: side-by-side LLM outputs.
• app/app.py: Streamlit app.

v0.2.0

This release adds a full Retrieval-Augmented Generation (RAG) pipeline over the Amazon Books dataset.

New features

• RAG pipeline (src/rag_pipeline.py): chains hybrid retrieval, context building, prompt formatting, and LLM generation into a single invoke() call
• Hybrid retriever (src/hybrid.py): combines BM25 and semantic search using round-robin interleaving, returning results with their source (BM25, Semantic, or BM25 + Semantic)
• LLM integration: Groq API with LLaMA 3.2 90B and automatic model fallback; SimpleLLM placeholder for demo mode
• Prompt templates (src/prompts.py): Balanced and Strict variants, selectable in the app
• Document chunker (src/chunking.py): 500-char chunks with 50-char overlap for long reviews
• Updated Streamlit app: four tabs (BM25, Semantic, Hybrid, RAG); RAG tab shows generated answer above source documents

Notebooks

• notebooks/06_hybrid_retrieval.ipynb: hybrid retrieval with auto-generation of src/hybrid.py
• notebooks/07_rag_pipeline.ipynb: full RAG pipeline with evaluation on 10 queries

Evaluation

• Qualitative evaluation of 5 queries (Easy / Medium / Complex) documented in results/milestone2_discussion.md.

v0.1.0

Milestone 1: Dual-Method Retrieval System

What's Included

• BM25 Keyword Search: Fast, exact-match retrieval with consistent tokenization
• Semantic Search: Embedding-based retrieval using all-MiniLM-L6-v2 + FAISS (IndexFlatL2)
• Hybrid Search: Combined BM25 + Semantic with adjustable weights
• Web App: Streamlit interface with query input, method selection, results display
• Evaluation: 10 diverse queries (4 Easy, 3 Medium, 3 Complex) with detailed comparisons
• Complete Documentation: Comprehensive README, data processing pipeline, preprocessing steps

Key Features

• Consistent text preprocessing (lowercase, remove punctuation, normalize whitespace)
• Top-5 results with titles, review previews, ratings, scores
• 2800 word README with full reproducibility instructions
• Professional evaluation with strengths/weaknesses analysis
• BM25 query speed under 10ms, Semantic search around 100ms

Files

• src/bm25.py - BM25Retriever class
• src/semantic.py - SemanticRetriever class
• src/utils.py - Preprocessing and tokenization
• notebooks/01_exploration.ipynb - Data exploration
• notebooks/02-05 - Data processing and indexing
• app/app.py - Streamlit application
• results/milestone1_discussion.md - Evaluation with 10 queries and 5 detailed comparisons